Authors

Abstract

The diamondback moth (Plutella xylostella) is a destructive pest that severely compromises Chinese cabbage production. Infestations caused by this pest significantly reduce both yield and quality, making efficient and accurate detection crucial for cultivation management. To address the challenges of detecting small targets and extracting phenotypic features in complex environments, this study proposes SAFF-YOLO—a YOLO11-based pest detection algorithm specifically designed for diamondback moths in Chinese cabbage fields. First, the loss function was refined to enhance the model's learning capacity for pest samples, optimizing it for precision-driven bounding box regression. Second, Alterable Kernel Convolution (AKConv) was incorporated into the backbone network, strengthening feature extraction capabilities while reducing model parameters. Third, Single-Head Self-Attention (SHSA) was integrated into the C2PSA (Channel and Position Spatial Attention) module, enhancing the backbone network's feature processing efficacy. Fourth, the neck network employed Frequency-aware Feature Fusion (FreqFusion) as the upsampling operator, specifically designed for precise localization of densely distributed targets. Finally, the Feature Auxiliary Fusion Single-Stage Head (FASFFHead) detection module was implemented to boost multi-scale target detection adaptability. Experimental results demonstrate that SAFF-YOLO achieved detection metrics of 90.7% precision, 89.4% recall, and 92.4% mAP50 for diamondback moths in Chinese cabbage, representing improvements of 7.4%, 8.0%, and 8.4% respectively over YOLO11. With only 7.3 million parameters and computational cost of 12.8 GFLOPs, this corresponds to 60.1% and 40.7% reductions compared to the baseline model. These results confirm an optimal balance between model lightweighting and high detection accuracy. Under complex field conditions characterized by small and densely distributed targets, severe background interference, and intense illumination, SAFF-YOLO consistently demonstrates robust detection capabilities, effectively reducing both false negative and false positive rates while maintaining high operational robustness. This research provides a practical solution for real-time diamondback moth detection in field-grown Chinese cabbage.

Abstract in Chinese

小菜蛾是严重危害白菜生产的害虫，其导致的虫害会使白菜产量、质量严重下降，因此高效、准确地检测小菜蛾对白菜栽培至关重要。针对复杂环境下小菜蛾检测存在目标小、表型特征提取困难等问题，本研究提出了基于YOLO11的白菜小菜蛾害虫检测算法SAFF-YOLO。首先，改进损失函数来增强模型对害虫样本的学习能力，使其更适合边界框回归的准确性需求；引入可变核卷积（Alterable Kernel Convolution，AKConv）作为主干网络，增强了特征提取能力，减少了模型参数的数量；将单头自注意力（Single-Head Self-Attention，SHSA）集成至C2PSA（Channel and Position Spatial Attention）模块中，提高了骨干网络的特征处理能力；颈部网络使用频率感知特征融合（Frequency-aware Feature Fusion，FreqFusion）作为上采样算子，旨在更好的对密集目标识别定位；最后通过FASFFHead（Feature Auxiliary Fusion Single-Stage Head）检测头增强模型对不同尺度目标的检测能力。试验结果表明，SAFF-YOLO对白菜小菜蛾的检测准确率、召回率、平均精度均值（mean average precision，mAP50）达到90.7%、89.4%和92.4%，对比YOLO11各提高了7.4%、8.0%和8.4%，且参数量为7.3Ｍ，每秒浮点运算次数（Giga Floating-point Operations Per Second，GFLOPs）为12.8，相较于基准模型分别降低60.1%和40.7%，实现了模型轻量化和较高检测精度的平衡。在小菜蛾小且密集、背景干扰严重、光照强烈等复杂环境下，SAFF-YOLO均能较好地识别出目标个体，有效地降低漏检率和误检率，具有较好的鲁棒性。本研究可为田间白菜小菜蛾实时检测提供有效技术支持。